Assessing reliability of BERT-based models on question answering tasks

Cross-Domain Faithfulness Evaluation of SHAP and Attention-Based Explanations in Transformer NLP Models

Transformer-based models such as BERT, RoBERTa, DistilBERT, and DeBERTa have achieved remarkable performance across a wide range of natural language processing (NLP) tasks. However, their decision-making processes remain difficult to interpret, particularly in high-risk applications such as hate speech detection, where unreliable explanations may undermine model transparency, trust, and accountability. This study investigates whether explainability methods remain faithful and stable under domain shift in transformer-based text classification. Four transformer architectures were fine-tuned and evaluated on two linguistically distinct datasets: IMDb Movie Reviews and Hate Speech Offensive. Model performance and explanation quality were assessed using classification accuracy, macro F1-score, top-k token-removal faithfulness analysis, and cross-domain Spearman rank correlation. Experimental results show that DeBERTa achieved the highest classification performance, reaching accuracies of 95.6% on IMDb and 91.3% on Hate Speech. Across all evaluated models and datasets, SHAP consistently produced higher faithfulness scores than attention-based explanations. Cross-domain analysis further revealed reduced agreement between SHAP and attention-based explanations under domain shift, indicating lower explanation consistency across linguistically distinct domains. Qualitative error analysis further showed that implicit sentiment, sarcasm, and domain-specific slang remain major sources of prediction errors. Overall, the results demonstrate that superior predictive performance does not necessarily correspond to higher explanation faithfulness or stronger cross-domain stability. These findings highlight the importance of jointly evaluating predictive performance, explanation faithfulness, and explanation robustness when developing trustworthy transformer-based NLP systems.

A statistical framework for reliability evaluation of large language models

The growing use of Large Language Models (LLMs) in real-world applications has increased the need for reliable evaluation methods. This study proposes a statistical framework for evaluating LLM reliability by jointly considering model accuracy, hallucination rate, and cross-domain performance stability. Based on the TruthfulQA dataset, 300 test questions were selected using a stratified sampling strategy, and responses were generated by DeepSeek-V4-Flash and GPT-4o-mini. GPT-4o was then employed as an automated evaluator to assess the 600 generated responses. Model reliability was evaluated using three dimensions: Accuracy, Hallucination Rate, and Cross-Domain Stability. In particular, the proposed CDS metric quantifies the consistency of model performance across different knowledge categories. A Weighted Geometric Mean was further employed to construct an overall reliability score. The experimental results show that the proposed framework can identify reliability differences between different LLMs. DeepSeek-V4-Flash achieved an overall reliability score of 0.770, compared with 0.678 for GPT-4o-mini. Sensitivity analysis further shows that the model ranking remains unchanged under different weight configurations, indicating that the proposed framework is reasonably robust to variations in indicator weights. This study provides a statistical framework for quantitatively assessing LLM reliability under multiple evaluation criteria.

Understanding the gap in evaluation frameworks and metrics for RAG-based question-answering system

The Pitfalls of Maximum Likelihood Training of TPMs for Controllable Language Generation

This work finds that likelihood-trained TPMs can result in failed generations due to overly large corrections to the LM’s logits, and trains TPMs with LM-aligned objectives that better align with the LM token-probability space.

Knowing Before Answering: Decoding Language Models for Reliable RAG

The results suggest that language models internally encode whether retrieved evidence is sufficient to support answering, and that this signal can be decoded reliably for RAG triage.

Improving Trustworthiness in Visual Question Answering Via Question-Conditioned Cross-Modal Verification

Visual Question Answering (VQA) is a challenging cross-disciplinary task that combine natural language processing and computer vision together. VQA answer natural language questions based on images. Current VQA systems prone to generate plausible but incorrect responses without indicating uncertainty. To address this limitation, we propose a training-free cross-modal answer verification framework. This framework based on question-conditioned CLIP scoring as a post-hoc reliability estimator for vision-language models. The technique improves the ability to distinguish between right and wrong short-form VQA answers by evaluating candidate solutions jointly with the original question using type-aware distractor pools. Experiments on the VQAv2 validation set evaluate raw accuracy, verified accuracy, coverage, and accuracy gain under threshold-based selective prediction. Results shows consistent reliability improvements across diverse model architectures with gains ranging from +1.13 to +32.98 percentage points depending on coverage and model characteristics. The proposed framework is model-agnostic, requires no retraining or architectural modification, and improves the trustworthiness of multimodal systems by selectively filtering unreliable predictions. These findings show that question-conditioned CLIP verification provides an effective and scalable reliability layer for VQA systems.